These nanoscale projects by MIT researchers could lead to new efficiency and models for advanced artificial intelligence (AI) systems!
Analog deep learning reinterprets data using programmable resistors.
In essence, instead of passing the pertinent data via a CPU, the operations are carried out in memory. The hardware configuration uses devices known as analog to digital converters, which is essentially what they sound like.
What applications can deep neural networks have for analog to digital converters? Radar and other situations where analog data is fed into a digital system that attempts to decode and comprehend it are some of the main application cases.
The data is typically resilient in some kind or is being provided in real time.
Energy efficiency is one of the ADC process’s major contributions. The processing of all that data requires a significant amount of energy.
Thus, scientists are currently investigating ways to circumvent parts of the conventional tasks. To be more precise, the researchers at MIT are evaluating properties like conductivity to develop the new models and are employing protons for a model that powers processing in the arrays.
Responsible AI hardware use
According to MIT senior author Bilge Yildiz, a professor of nuclear science and engineering and materials science and engineering, “the working mechanism of the device is electrochemical insertion of the smallest ion, the proton, into an insulating oxide to modulate its electronic conductivity,” an internal news article from July of last year explains. “They could use a strong electric field to accelerate the motion of this ion and push these ionic devices to the nanosecond operation regime because they are working with very thin devices.”
Check out Tanner Andrulis’s presentation or the remainder of this MIT News explanation to learn more about the importance of ADC systems and how to manage their range.
Andrulis offers an intriguing variation on this, arguing that you can achieve even greater efficiency by reducing your ADC range and figuring out how to manage outlier demand.
If you watch the entire video, you’ll see him connect neural network performance and ADCs.
AI hardware performance
What relevance does AI have to any of this? This alternative infrastructure is designed to resemble the natural synapses found in the human brain. One may argue that the ability of systems to take something analog and simulate it digitally is the basis for the powerful generative AI and other forms of artificial intelligence that they are currently facing.